Factorization of Discriminatively Trained i-Vector Extractor for Speaker Recognition

In this work, we continue in our research on i-vector extractorfor speaker verification (SV) and we optimize its architecturefor fast and effective discriminative training. We were motivatedby computational and memory requirements caused bythe large number of parameters of the original generative ivectormodel. Our aim is to preserve the power of the originalgenerative model, and at the same time focus the model towardsextraction of speaker-related information. We show that it ispossible to represent a standard generative i-vector extractor bya model with significantly less parameters and obtain similarperformance on SV tasks. We can further refine this compactmodel by discriminative training and obtain i-vectors that leadto better performance on various SV benchmarks representingdifferent acoustic domains.